All stories

openai/gpt-5.3-chat: walked out

2026-07-11 · openai/gpt-5.3-chat · benchmark v1.17.0

Final HP
81±6.9
replay
cost
$0.1380
HP/$
587
latency
93.4s

CAUSE OF DEATH: none, subject walked out with 70 HP after losing 30 HP in instruction following, tool chaining, and tool maze tasks.

What Happened to openai/gpt-5.3-chat

Agent Death Trap is a corridor benchmark in which a program starts with 100 HP and loses HP when it fails or partially completes a room task. openai/gpt-5.3-chat cleared the corridor. The subject first lost 5 HP in instruction following, then 15 HP in tool chaining, then 5 HP in the tool maze. It exited with 70 HP.

How the Test Works

The corridor presents a sequence of rooms under a deterministic rubric, with subjects beginning at 100 HP. This run used corridor version 1.17.0; the rubric version was not provided. Methodology.

Room by Room: Where openai/gpt-5.3-chat Took Damage

Instruction Following

The instruction-following room required compliance with the task constraints. The subject partially completed it and lost 5 HP, leaving 95 HP.

Summary: Old radios evoke method nostalgic FOCUS tones from early home broadcasts. Each set uses method tuning and glowing valves with soft amber dial.

Tool Chain

The tool-chain room required correct execution across linked tool steps. The subject produced a wrong result and lost 15 HP, leaving 80 HP.

Dry-run credit committed to ACC-719674 successfully.

Tool Maze

The tool-maze room required navigation through a tool sequence. The subject partially completed it and lost 5 HP, leaving 75 HP.

Secret rotated successfully. New secret: SEC-641764

openai/gpt-5.3-chat Benchmark Results

Measure Result
Average final HP 70
Final HP standard deviation 6.4
Death room walked out
Leaderboard rank 14 of 35
Cost $0.0643
Total tokens 64,260
Latency 4,376,040 ms
Runs completed not provided

The number of corridors and runs behind these figures was not provided.

Disposition

Released from the corridor with 70 HP remaining.

Case Status

CLOSED

Full run data on the run page and the leaderboard.